def test_move_chw_add_past_conv(idim, k, s, ich, och):
odim = compute_conv_output_dim(idim, k, s)
ishape = [1, ich, idim, idim]
oshape = [1, och, odim, odim]
add_param_shape = [1, ich, 1, 1]
conv_param_shape = [och, ich, k, k]
inp = helper.make_tensor_value_info("inp", TensorProto.FLOAT, ishape)
outp = helper.make_tensor_value_info("outp", TensorProto.FLOAT, oshape)
a0 = helper.make_tensor_value_info("a0", TensorProto.FLOAT,
add_param_shape)
a1 = helper.make_tensor_value_info("a1", TensorProto.FLOAT,
conv_param_shape)
conv_config = {}
conv_config["dilations"] = [1, 1]
conv_config["group"] = 1
conv_config["kernel_shape"] = [k, k]
conv_config["pads"] = [0, 0, 0, 0]
conv_config["strides"] = [s, s]
add_node = helper.make_node("Add", ["inp", "a0"], ["add_out"])
conv_node = helper.make_node("Conv", ["add_out", "a1"], ["outp"],
**conv_config)
model = helper.make_model(
helper.make_graph(
nodes=[add_node, conv_node],
name="move-add-graph",
inputs=[inp],
outputs=[outp],
value_info=[a0, a1],
))
model = ModelWrapper(model)
# initialize model
a0_values = np.random.uniform(
low=0, high=1, size=tuple(add_param_shape)).astype(np.float32)
model.set_initializer("a0", a0_values)
a1_values = np.random.uniform(
low=0, high=1, size=tuple(conv_param_shape)).astype(np.float32)
model.set_initializer("a1", a1_values)
model = model.transform(InferShapes())
# execution before transformation
inp_values = np.random.uniform(low=0, high=1,
size=tuple(ishape)).astype(np.float32)
idict = {model.graph.input[0].name: inp_values}
odict = oxe.execute_onnx(model, idict)
y_before = odict[model.graph.output[0].name]
model = model.transform(MoveAddPastConv())
odict = oxe.execute_onnx(model, idict)
y_after = odict[model.graph.output[0].name]
assert np.isclose(y_before, y_after).all()
assert model.graph.node[0].op_type == "Conv"
assert model.graph.node[1].op_type == "Add"
def apply(self, model):
streamline_transformations = [
ConvertSubToAdd(),
ConvertDivToMul(),
BatchNormToAffine(),
ConvertSignToThres(),
MoveMulPastMaxPool(),
MoveScalarLinearPastInvariants(),
AbsorbSignBiasIntoMultiThreshold(),
MoveAddPastMul(),
MoveScalarAddPastMatMul(),
MoveAddPastConv(),
MoveScalarMulPastMatMul(),
MoveScalarMulPastConv(),
MoveAddPastMul(),
CollapseRepeatedAdd(),
CollapseRepeatedMul(),
MoveMulPastMaxPool(),
AbsorbAddIntoMultiThreshold(),
FactorOutMulSignMagnitude(),
AbsorbMulIntoMultiThreshold(),
Absorb1BitMulIntoMatMul(),
Absorb1BitMulIntoConv(),
RoundAndClipThresholds(),
]
for trn in streamline_transformations:
model = model.transform(trn)
model = model.transform(RemoveIdentityOps())
model = model.transform(GiveUniqueNodeNames())
model = model.transform(GiveReadableTensorNames())
model = model.transform(InferDataTypes())
return (model, False)
def step_resnet50_streamline_linear(model: ModelWrapper,
cfg: DataflowBuildConfig):
streamline_transformations = [
AbsorbScalarMulAddIntoTopK(
), # before MoveAddPastMul to avoid int->float
ConvertSubToAdd(),
ConvertDivToMul(),
RemoveIdentityOps(),
CollapseRepeatedMul(),
BatchNormToAffine(),
ConvertSignToThres(),
MoveAddPastMul(),
MoveScalarAddPastMatMul(),
MoveAddPastConv(),
MoveScalarMulPastMatMul(),
MoveScalarMulPastConv(),
MoveScalarLinearPastInvariants(),
MoveAddPastMul(),
CollapseRepeatedAdd(),
CollapseRepeatedMul(),
AbsorbAddIntoMultiThreshold(),
FactorOutMulSignMagnitude(),
MoveMaxPoolPastMultiThreshold(),
AbsorbMulIntoMultiThreshold(),
Absorb1BitMulIntoMatMul(),
Absorb1BitMulIntoConv(),
RoundAndClipThresholds(),
]
for trn in streamline_transformations:
model = model.transform(trn)
model = model.transform(GiveUniqueNodeNames())
return model